We use cookies to analyze the browsing and usage of our website and to personalize your experience. You can disable these technologies at any time, but this may limit certain functionalities of the site. Read our Privacy Policy for more information.
Setting cookies
You can enable and disable the types of cookies you wish to accept. However certain choices you make could affect the services offered on our sites (e.g. suggestions, personalised ads, etc.).
Essential cookies
These cookies are necessary for the operation of the site and cannot be deactivated. (Still active)
Analytics cookies
Do you accept the use of cookies to measure the audience of our sites?
Multimedia Player
Do you accept the use of cookies to display and allow you to watch the video content hosted by our partners (YouTube, etc.)?
Publications
Beyond the Safety Bundle: Auditing the Helpful and Harmless Dataset
In an effort to mitigate the harms of large language models (LLMs), learning from human feedback (LHF) has been used to steer LLMs towards o… (see more)utputs that are intended to be both less harmful and more helpful. Despite the widespread adoption of LHF in practice, the quality of this feedback and its effectiveness as a safety mitigation technique remain unclear. This study addresses these issues by auditing the widely-used Helpful and Harmless (HH) dataset by Anthropic. Our work includes: (1) a thorough investigation of the dataset's content through both manual and automated evaluation; (2) experiments demonstrating the dataset's impact on models' safety; and (3) an analysis of the 100 most influential papers citing this dataset. Through our audit, we showcase how conceptualization failures and quality issues identified in the HH dataset can create additional harms by leading to disparate safety behaviors across demographic groups. Our findings highlight the need for more nuanced, context-sensitive approaches to safety mitigation in LLMs.
Over the past decade, Deep Learning (DL) has become an integral part of our daily lives. This surge in DL usage has heightened the need for … (see more)developing reliable DL software systems. Given that fault localization is a critical task in reliability assessment, researchers have proposed several fault localization techniques for DL-based software, primarily focusing on faults within the DL model. While the DL model is central to DL components, there are other elements that significantly impact the performance of DL components. As a result, fault localization methods that concentrate solely on the DL model overlook a large portion of the system. To address this, we introduce FL4Deep, a system-level fault localization approach considering the entire DL development pipeline to effectively localize faults across the DL-based systems. In an evaluation using 100 faulty DL scripts, FL4Deep outperformed four previous approaches in terms of accuracy for three out of six DL-related faults, including issues related to data (84%), mismatched libraries between training and deployment (100%), and loss function (69%). Additionally, FL4Deep demonstrated superior precision and recall in fault localization for five categories of faults including three mentioned fault types in terms of accuracy, plus insufficient training iteration and activation function.
Speech impairments in Parkinson's disease (PD) provide significant early indicators for diagnosis. While models for speech-based PD detectio… (see more)n have shown strong performance, their interpretability remains underexplored. This study systematically evaluates several explainability methods to identify PD-specific speech features, aiming to support the development of accurate, interpretable models for clinical decision-making in PD diagnosis and monitoring. Our methodology involves (i) obtaining attributions and saliency maps using mainstream interpretability techniques, (ii) quantitatively evaluating the faithfulness of these maps and their combinations obtained via union and intersection through a range of established metrics, and (iii) assessing the information conveyed by the saliency maps for PD detection from an auxiliary classifier. Our results reveal that, while explanations are aligned with the classifier, they often fail to provide valuable information for domain experts.
There is a growing interest in training domain-expert LLMs that excel in specific technical fields compared to their general-purpose instruc… (see more)tion-tuned counterparts. However, these expert models often experience a loss in their safety abilities in the process, making them capable of generating harmful content. As a solution, we introduce an efficient and effective merging-based alignment method called \textsc{MergeAlign} that interpolates the domain and alignment vectors, creating safer domain-specific models while preserving their utility. We apply \textsc{MergeAlign} on Llama3 variants that are experts in medicine and finance, obtaining substantial alignment improvements with minimal to no degradation on domain-specific benchmarks. We study the impact of model merging through model similarity metrics and contributions of individual models being merged. We hope our findings open new research avenues and inspire more efficient development of safe expert LLMs.
There is a growing interest in training domain-expert LLMs that excel in specific technical fields compared to their general-purpose instruc… (see more)tion-tuned counterparts. However, these expert models often experience a loss in their safety abilities in the process, making them capable of generating harmful content. As a solution, we introduce an efficient and effective merging-based alignment method called \textsc{MergeAlign} that interpolates the domain and alignment vectors, creating safer domain-specific models while preserving their utility. We apply \textsc{MergeAlign} on Llama3 variants that are experts in medicine and finance, obtaining substantial alignment improvements with minimal to no degradation on domain-specific benchmarks. We study the impact of model merging through model similarity metrics and contributions of individual models being merged. We hope our findings open new research avenues and inspire more efficient development of safe expert LLMs.
A key objective of interpretability research on large language models (LLMs) is to develop methods for robustly steering models toward desir… (see more)ed behaviors. To this end, two distinct approaches to interpretability -- ``bottom-up"and ``top-down"-- have been presented, but there has been little quantitative comparison between them. We present a case study comparing the effectiveness of representative vector steering methods from each branch: function vectors (FV; arXiv:2310.15213), as a bottom-up method, and in-context vectors (ICV; arXiv:2311.06668) as a top-down method. While both aim to capture compact representations of broad in-context learning tasks, we find they are effective only on specific types of tasks: ICVs outperform FVs in behavioral shifting, whereas FVs excel in tasks requiring more precision. We discuss the implications for future evaluations of steering methods and for further research into top-down and bottom-up steering given these findings.
Understanding the mechanisms through which neural networks extract statistics from input-label pairs through feature learning is one of the … (see more)most important unsolved problems in supervised learning. Prior works demonstrated that the gram matrices of the weights (the neural feature matrices, NFM) and the average gradient outer products (AGOP) become correlated during training, in a statement known as the neural feature ansatz (NFA). Through the NFA, the authors introduce mapping with the AGOP as a general mechanism for neural feature learning. However, these works do not provide a theoretical explanation for this correlation or its origins. In this work, we further clarify the nature of this correlation, and explain its emergence. We show that this correlation is equivalent to alignment between the left singular structure of the weight matrices and the newly defined pre-activation tangent features at each layer. We further establish that the alignment is driven by the interaction of weight changes induced by SGD with the pre-activation features, and analyze the resulting dynamics analytically at early times in terms of simple statistics of the inputs and labels. We prove the derivative alignment occurs with high probability in specific high dimensional settings. Finally, motivated by the observation that the NFA is driven by this centered correlation, we introduce a simple optimization rule that dramatically increases the NFA correlations at any given layer and improves the quality of features learned.
We conduct a large-scale empirical user study in a live setup to evaluate the acceptance of LLM-generated comments and their impact on the r… (see more)eview process. This user study was performed in two organizations, Mozilla (which has its codebase available as open source) and Ubisoft (fully closed-source). Inside their usual review environment, participants were given access to RevMate, an LLM-based assistive tool suggesting generated review comments using an off-the-shelf LLM with Retrieval Augmented Generation to provide extra code and review context, combined with LLM-as-a-Judge, to auto-evaluate the generated comments and discard irrelevant cases. Based on more than 587 patch reviews provided by RevMate, we observed that 8.1% and 7.2%, respectively, of LLM-generated comments were accepted by reviewers in each organization, while 14.6% and 20.5% other comments were still marked as valuable as review or development tips. Refactoring-related comments are more likely to be accepted than Functional comments (18.2% and 18.6% compared to 4.8% and 5.2%). The extra time spent by reviewers to inspect generated comments or edit accepted ones (36/119), yielding an overall median of 43s per patch, is reasonable. The accepted generated comments are as likely to yield future revisions of the revised patch as human-written comments (74% vs 73% at chunk-level).
In inverse reinforcement learning (IRL), an agent seeks to replicate expert demonstrations through interactions with the environment. Tradit… (see more)ionally, IRL is treated as an adversarial game, where an adversary searches over reward models, and a learner optimizes the reward through repeated RL procedures. This game-solving approach is both computationally expensive and difficult to stabilize. In this work, we propose a novel approach to IRL by direct policy optimization: exploiting a linear factorization of the return as the inner product of successor features and a reward vector, we design an IRL algorithm by policy gradient descent on the gap between the learner and expert features. Our non-adversarial method does not require learning a reward function and can be solved seamlessly with existing actor-critic RL algorithms. Remarkably, our approach works in state-only settings without expert action labels, a setting which behavior cloning (BC) cannot solve. Empirical results demonstrate that our method learns from as few as a single expert demonstration and achieves improved performance on various control tasks.
Neuronal network activity is thought to be structured around the activation of assemblies, or low-dimensional manifolds describing states of… (see more) activity. Both views describe neurons acting not independently, but in concert, likely facilitated by strong recurrent excitation between them. The role of inhibition in these frameworks – if considered at all – is often reduced to blanket inhibition with no specificity with respect to which excitatory neurons are targeted. We analyzed the structure of excitation and inhibition in the MICrONS 1mm3 dataset, an electron microscopic reconstruction of a piece of cortical tissue. We found that excitation was structured around a feed-forward flow in non-random motifs of seven or more neurons. This revealed a structure of information flow from a small number of sources to a larger number of potential targets that became only visible when larger motifs were considered instead of individual pairs. Inhibitory neurons targeted and were targeted by neurons in specific sequential positions of these motifs. Additionally, disynaptic inhibition was strongest between target motifs excited by the same group of source neurons, implying competition between them. The structure of this inhibition was also highly specific and symmetrical, contradicting the idea of non-specific blanket inhibition. None of these trends are detectable in only pairwise connectivity, demonstrating that inhibition is specifically structured by these large motifs. Further, we found that these motifs represent higher order connectivity patterns which are present, but to a lesser extent in a recently released, detailed computational model, and not at all in a distance-dependent control. These findings have important implications for how synaptic plasticity reorganizes neocortical connectivity to implement learning and for the specific role of inhibition in this process.
